SIS-CAM: An Interpretability Analysis Method for the Security of Convolutional Neural Network Models Based on Image Big Data

The opacity of deep Convolutional Neural Network (CNN) models presents significant challenges to their interpretability and security, hindering understanding of their internal decision-making processes. To address these challenges, this paper introduces SIS-CAM, a novel interpretability method that enhances CNN transparency and security by improving saliency map visualization. SIS-CAM optimizes gradient calculations by squaring gradient values during backpropagation and fuses initial saliency maps with input images through feature fusion. It iteratively refines these maps by applying masking techniques, averaging, and linearly combining them with the initial saliency maps, reducing noise and improving clarity. Additionally, SIS-CAM generates accurate masks to highlight key boundary features and significant neuron activations through enhanced gradient computations. The method’s effectiveness is rigorously evaluated on the ILSVRC2012val dataset, using both qualitative and quantitative metrics such as Deletion and Insertion tests, Average Drop, Average Insertion scores, Class Discriminative Visualization, and sanity checks. Experimental results demonstrate that SIS-CAM effectively reduces noise, identifies critical features, and outperforms existing methods in visualizing CNN decision-making, making it a valuable tool for improving model interpretability and ensuring security, particularly in high-stakes applications involving large-scale image data.